Vega-U (Qwen3.5-0.8B reader)
Data: JevBench v1.6.39, published 2026-10-22
Capability 36.5 · Not ranked in the current release
latency 2.5× Jev (v1.5 reference)
Last measured: 2026-10-18 · measurement release: v1.6.29
vega-U checkpoint apache-2.0 (card and metadata agree); vegaml package apache-2.0; frozen reader Qwen/Qwen3.5-0.8B apache-2.0
intelligence
5.5
calibration
67.5
speed
73.8
cost
79.2
Composite (secondary)
0.2 · Composite rank #168
Cost and measurement conditions
$0.0049 per 1,000 decisions (estimated)
ESTIMATE (*est.) through the frozen pricing_v15 path (module sha256 6202ef78...): vega-U is a 14.3 M decision engine on a FROZEN Qwen/Qwen3.5-0.8B reader, which resolves to deepinfra:Qwen/Qwen3.5-0.8B in the dated 25 Sep DeepInfra catalog snapshot (sha256 84eb7926acd4e93b71d695be24898e2ca999b20011d1b66cc0348ce7c6c3546f), USD 0.01/M input and USD 0.05/M output - the same basis and the same numbers the live vegaml-0.8b row carries. The model generates no text: 0 output tokens on every item, so cost = input tokens x USD 0.01/M. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.004930 per 1,000 decisions (0.076x the official cost cap USD 0.064595), against USD 0.012253 over all 1,500 items (v1.6.29).
Measured p50 latency: 0.695 s (on-CPU, no network, loopback; not directly comparable with API latency, which we time end to end from our measuring host, network included)
Ranking and cap use: 1.54 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.695 s on-CPU, no network (timed on our CPU host over loopback). Not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and latency cap use the same estimate for every self-hosted row: 2 × measured + 150 ms = 1.54 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
our own CPU server (EPYC 7502P, 32 of 64 threads, 96 GB), disposable docker container with --network none (cut proven from inside before the sealed mount), --cap-drop=ALL, no-new-privileges, unprivileged user, one read-only sealed input mount; reviewed vegaml==0.12.2, torch 2.14.1+cpu, transformers 5.19.0; fp32 throughout (the author's own backbone_dtype() choice on CPU); context 73,728, and a longer state is cut on the right by the author's own build_prompt - 15 of the 1,500 items sit at that cap and are answered on a truncated state. Reference PyTorch paths only.